Storing is not answering
Nora · 0:00I want you to imagine something for a second. Think about your own digital life, like the unread PDFs just piling up in your downloads folder.
Miles · 0:08Oh, yeah. Hundreds of them.
Nora · 0:09Right. Or those 50-page terms of service you blindly accept, the dense legal contracts, instruction manuals. Imagine a world where you never have to actually read a single one of them ever again.
Miles · 0:20Which sounds amazing.
Nora · 0:21It does. You just drop this massive digital file into a folder and you instantly know everything inside of it. simply because you can ask it questions.
Miles · 0:31Yeah, and I mean, it fundamentally changes the definition of literacy at that point. You know, you're no longer extracting information. You're interrogating a system that has already done the reading for you.
Nora · 0:40Exactly. And this isn't, it's not some distant science fiction concept. We are doing a deep dive today into a massive shift that's happening right under our feet. The end of human-centric reading.
Miles · 0:53Yeah, it's happening fast.
Nora · 0:54Really fast. We're looking at a pretty fascinating stack of sources today. technical papers, speeches from tech leaders at companies like, you know, OpenAI and Microsoft, plus some new product architectures. And they all point to this same conclusion. Machines are taking over the consumption of text.
Miles · 1:12Which is a huge paradigm shift.
Nora · 1:14It really is. Okay, let's unpack this. Because to understand how our own reading habits are about to change, we first have to look at the invisible plumbing of the internet itself. The very content we consume is being fundamentally redesigned. Right.
Miles · 1:27Right. Because, well, the Internet, as you and I know it, it's basically an artifact of human biology.
Nora · 1:33How do you mean?
Miles · 1:34Well, for the last 30 years, we've built the web entirely around the limitations and, you know, preferences of the human eyeball.
Nora · 1:40Oh, sure.
Miles · 1:40We use flashy graphics, nested drop-down menus, cascading style sheets, clickbait headlines, all of it engineered to capture human attention. But, you know, AI researcher Andrej Karpathy recently put forward this thesis that this era is practically over.
Nora · 1:55Wow. Over completely.
Miles · 1:56Yeah. He argues that very soon, like 99.9% of all content optimization is going to be focused on AI, not humans.
Nora · 2:04Which, I mean, that makes perfect sense when you look at how hostile the current Internet is to a machine.
Miles · 2:09Oh, it's a nightmare for them.
Nora · 2:10Right. If an AI tries to read a standard web page, it has to wade through this swamp of HTML tags and formatting code and pop-up scripts just to find the actual information.
Miles · 2:20Exactly. And Carpenter points out that the standard way we build software documentation like. creating these beautiful interconnected web pages, it's becoming entirely obsolete.
Nora · 2:30Because the AI language model just doesn't care about your beautiful user interface.
Miles · 2:34No, it doesn't care at all. It just wants the raw data. So Karpathy suggests that instead of building a complex website, modern projects should just provide a single plain text markdown file.
Nora · 2:45A .md file.
Miles · 2:46Right, markdown, which strips away all the visual styling. It just uses really simple simple text symbols to denote headings and lists. It's highly structured, incredibly lightweight, and it's explicitly designed to be dropped straight into an AI's context window. You
The agent-readable web
Nora · 3:02bypass the visual web entirely. And there's actually a brand new proposed web standard trying to formalize this bypass. It's called LE-TXT. Developed by Jeremy Howard. Right. And companies like Anthropic are already adopting it. You can think of it like, well, it's like a secret door built into a website site just for the bots.
Miles · 3:20That's a great way to put it.
Nora · 3:21Yeah. So when an agent visits a domain, instead of trying to parse the pretty human facing HTML, it just looks for this Elms dot TXT file. And that gives the AI a map directly to the clean machine readable markdown versions of the site.
Miles · 3:35And that concept of bypassing the human interface entirely, it's starting to collapse traditional software, too.
Nora · 3:42Oh, yeah. Microsoft CEO Satya Nadella was talking about this, right?
Miles · 3:45Yeah. He recently predicted that this new agentic AI era is going to totally dismantle traditional software as a service.
Nora · 3:53Suss. Which is just a wild prediction coming from the guy whose company basically invented modern enterprise software.
Miles · 3:59Right. He actually asked the question, you know, hey, why do I need Excel?
Nora · 4:04Which sounds completely absurd.
Miles · 4:06It does, until you look at the mechanics of what most software actually is. Nadella points out that the vast majority of business applications are just CRUD databases.
Nora · 4:16CRUD, meaning create, read, update, delete.
Miles · 4:20Exactly. They are just places to store and change data wrapped in a layer of business logic and topped with a graphical user interface so, you know, a human can click around. Right. But if you have an AI agent that can understand the business logic natively, the interface becomes totally useless. You don't need a dashboard with 100 buttons.
Nora · 4:39You just tell an AI assistant to analyze the data using Python and it executes the code directly against the database.
Miles · 4:45Exactly. The entire visual layer of the software just evaporates.
Nora · 4:49It's like, I was thinking about this, building these complex graphical interfaces for an AIDU is essentially like putting a tuxedo on a dog.
Miles · 4:57A tuxedo on a dog.
Nora · 4:58Yeah. I mean, it might look cute to us, but it completely ignores how the dog actually navigates the world. The dog doesn't care about the bow tie. It just wants the treat.
Miles · 5:05That is very true.
Nora · 5:07The AI doesn't care about your dropdown menu. It just wants the raw text. But wait, I do have a pushback here from the sources.
Miles · 5:14Okay, let's hear it.
Nora · 5:15If the interface disappears and these AI companies are licensing the world's information to feed their models, aren't these language models becoming the ultimate dangerous gatekeepers? I mean, they get to decide what their systems see and, by extension, what we know.
Miles · 5:31Yeah. And tracing the economic incentives here reveals a massive problem because the entire digital value chain today runs on human attention.
Nora · 5:41Right. Ads and subscriptions.
Miles · 5:42Exactly. The Internet is free because you look at advertisements. But machines don't click on banner ads.
Nora · 5:48No, they don't.
Miles · 5:48They don't buy things impulsively because of a pop up. So if machines become the primary consumers of the Web, that entire revenue model breaks. breaks. The companies controlling the AI will dictate the diet of the models, which means they fundamentally control the flow of information.
Nora · 6:04In a post-human reading internet.
Miles · 6:06Precisely.
Nora · 6:07And this isn't just about enterprise software or web developers. This shift is coming directly for your personal life. Think about your own digital footprint, you know, your emails, texts, photos, notes.
Miles · 6:17The scale of how we process that personal data is just exploding.
Nora · 6:21Yeah. Let's look at Sam Altman's recent talk at Sequoia's AI Ascent. He shared a stat to illustrate the scale. And actually, before I share it, let's just quickly define a term for everyone. A token.
Tokens, and how differently generations use AI
Miles · 6:33Good idea.
Nora · 6:33Think of a token as a piece of a word, maybe three or four letters. It's the fundamental unit of data in AI processes when it reads. So a few years ago, Altman noted that OpenAI's absolute heaviest user burned through about 100,000 tokens a month.
Miles · 6:48Which seemed like a lot then.
Nora · 6:49Right. But today, 100,000 tokens is the per capita average. Wow. Yeah. The top users are heading 100 billion tokens a month.
Miles · 6:58That is literally a million-fold increase. And, you know, Altman has a very specific architectural vision for where all those billions of tokens are going. He calls it the platonic ideal for AI.
Nora · 7:09Right. He envisions this tiny, hyper-efficient reasoning model, but with context window of one trillion tokens.
Miles · 7:18Which is massive.
Nora · 7:19To put that in perspective, that window is large enough to contain your entire life.
Miles · 7:23Every email you've ever sent.
Nora · 7:25Every book you've ever read. Every recording of every meeting you've ever sat in. And the model never has to be retrained on your data. Your life just continuously appends to this massive open window.
Miles · 7:37So you'd essentially walk around with a digital entity that has read everything you've ever experienced and can reason over it in real time.
Nora · 7:44Exactly. Exactly.
Miles · 7:45Which perfectly explains the generational divide we're already seeing in how these tools are adopted.
Nora · 7:51Oh, the data on this is fascinating.
Miles · 7:53It is. The data shows that older generations tend to treat AI like a glorified search engine, right? They ask a question, get a fact, move on. But people in their 20s and 30s are using it as a life advisor. They're pasting in long text threads and asking for relationship or career advice.
Nora · 8:10And then college students are taking it a step further. They're using AI as a full blown operating system. They aren't just asking questions. They're building complex workflows, linking disparate files together, having the AI manage the friction of their daily tasks. They're basically already treating the AI like that whole life context window Altman is talking about.
Miles · 8:31Yeah. And his timeline for this is incredibly aggressive. He predicts 2025 is the year AI agents start doing actual autonomous work, specifically in coding.
Nora · 8:41And by 2027, this intelligence moves into the physical world through robotics.
Miles · 8:46So it's right around the corner.
Pushing back on token maxing
Nora · 8:47I do have to push back on this, though, this whole token maxing vision. Let's look at the incentives driving it. NVIDIA sells the highly expensive GPUs required to process these massive token loads. OpenAI sells the tokens themselves. So isn't this whole life context window just a giant self-serving sales pitch?
Miles · 9:06You mean because they profit directly from it?
Nora · 9:08Yeah. Yeah. Of course they want to paint a future where AI voraciously reads billions of your personal tokens every single day. They are the ones selling the shovels and the gold rush.
Miles · 9:16Well, it's crucial to be skeptical of the infrastructure providers, absolutely. But NVIDIA's CEO, Jensen Huang, offers a pretty compelling counterargument, and it's based on friction.
Nora · 9:26How so?
Miles · 9:27He argues that tokens are the new currency of the digital age, and AI agents are the ultimate consumers of that currency. Imagine an AI agent encounters a brand new, highly complex piece of proprietary hardware.
Nora · 9:41Okay.
Miles · 9:41In the old paradigm, a team of engineers would have to spend weeks training a new model to understand that hardware. But in the new paradigm, the agent just reads the thousand-page engineering manual in a fraction of a second, consumes the tokens, and becomes an expert instantly.
Nora · 9:57Skipping the training phase entirely.
Miles · 9:58Right. That instantaneous speed of adaptation is why they believe massive token consumption is just inevitable.
What breaks when summaries replace sources
Nora · 10:04Here's where it gets really interesting, though. If we actually do this, if you hand over your entire archive of personal and professional documents to an AI, we have to look at the mechanics of how it actually stores and remembers that knowledge.
Miles · 10:16Because the current popular method has a hidden flaw.
Nora · 10:18A potentially fatal flaw. Yeah.
Miles · 10:20This brings us to a fascinating personal experiment run by Andrej Karpathy and the very sharp critique of that experiment by a software engineer named Anand Lahoti.
Nora · 10:30So Karpathy decided to build his own AI knowledge base. He took about 100 articles he wanted to understand, roughly 400,000 words, but he didn't sit down and read them to write summaries.
Miles · 10:41No, he used a language model as a compiler.
Nora · 10:43Right. The AI ingested all his raw documents and autonomously generated a highly structured, backlinked, marked-down wiki. It wrote its own concept articles, synthesized the themes, and essentially built a web of knowledge out of the raw text.
Miles · 10:57And, I mean, for a single user with 100 articles, it's a brilliant productivity hack.
Nora · 11:01Oh, for sure.
Miles · 11:02But Anand Lahoti warns that scaling this architecture leads to something he calls knowledge-based poisoning.
Nora · 11:08Sounds ominous.
Miles · 11:09It is. When a language model authors content like those synthesized wiki pages, and you index that AI-authored content as truth in your database, you are introducing unverifiable information. Over time, that information slowly drifts away from reality.
Nora · 11:24Let's break down the mechanics of why that drift happens, because Lahoti gives a brilliant concrete example in the sources.
Miles · 11:31Yeah, the contract example.
Nora · 11:32Right. Imagine you feed a vendor contract into the system. The raw legally binding contract states that your payment terms are net 30 with a 2% discount if paid within 10 days.
Miles · 11:43And that 2% discount is a highly specific financially critical detail.
Nora · 11:48Exactly. But when the AI compiles a general concept article about your company's vendor finances, it summarizes that clause simply as net 30 terms with early payment discounts.
Miles · 12:00It drops the specific 2%.
Nora · 12:01And it drops the 10 day window. It isn't hallucinating or lying. It's just providing a slightly lossy summary. But here is the fatal flaw. Six months later, your CFO asks the AI, hey, what is our typical early payment discount? The AI searches your database and finds its own summary article first because that article is highly linked and optimized. It reads the phrase early payment discounts, but it no longer has the 2% figure in its immediate context. So it hedges or worse, it guesses based on industry averages it learned during its initial training.
Miles · 12:33And Lahoti defines this as the fundamental difference between right time synthesis and query time synthesis.
Nora · 12:39Right. Break that down for us.
The librarian who points back to the shelf
Miles · 12:41So Karpathy's wiki uses right time synthesis. The AI reads the raw data, writes a brand new persistent article, and that new article is treated as fact. Query time synthesis operates completely differently. In that model, the AI might extract tags or structural metadata, but it never authors new prose to store in the database.
Nora · 13:00Just not saving a summary.
Miles · 13:02Exactly. When you ask a question, the system is forced to go back to the immutable original raw documents and synthesize a fresh answer right then and there.
Nora · 13:09Wow. So write time synthesis is basically a high-tech game of telephone.
Miles · 13:13Pretty much.
Nora · 13:14The AI is whispering summaries to itself over and over, slowly laundering its own tiny omissions into accepted corporate truth.
Miles · 13:22Yeah.
Nora · 13:22But I have to challenge Lahoti's solution here. Query time synthesis, forcing the machine to read the original raw sources every single time you ask a question, that required significantly more computing power.
Miles · 13:35It does. It's slower and more expensive.
Nora · 13:38So won't human nature just dictate that we accept a slightly poisoned knowledge base? I mean, in tech, convenience, speed, and cost almost always win out over perfect accuracy.
Miles · 13:49Well, that tension between cost and accuracy is exactly what enterprise AI is wrestling with right now. Lahoda uses a metaphor that clarifies the stakes perfectly.
Nora · 13:57Okay.
Miles · 13:58You want an AI librarian who writes index cards that point you back to the original books on the shelf. You do not want an AI librarian who writes brand new books, summarizing the old ones, and then shells them next to the originals, treating them as equals.
Nora · 14:09That makes total sense.
Miles · 14:10At a personal scale, you know, you might be able to manually audit the drift, but across a massive organization with millions of documents, that drift is an existential threat. Eventually, nobody can tell the difference between the actual legally binding contract and the AI's degraded memory of the contract.
Nora · 14:30So what does this all mean for the actual tools we are going to use? Because this isn't just a theoretical debate between engineers.
Miles · 14:36No, not at all.
Nora · 14:37Startups are actively trying to productize this concept of read nothing, and they are making hard architectural choices right now.
Miles · 14:45Yeah, let's look at the product fact sheet for a new tool called DocuStrata.
Nora · 14:48Their tagline literally embodies the theme of our deep dive today. Read nothing. Know everything.
Miles · 14:55And their origin story perfectly illustrates the pain point they're solving. The founder had accumulated an archive of over 100,000 personal documents in Evernote.
Nora · 15:04That's a lot of notes.
Cheap storage, expensive answers
Miles · 15:05Yeah. And whenever they dramatically hiked their subscription prices, the founder realized something profound. They had never actually read most of those documents and they probably never would. So instead of migrating to a cheaper digital filing cabinet, they decided to build a machine that could read the archive for them.
Nora · 15:22And it is specifically designed to ingest all those messy pre-LLM formats that we talked about earlier. Old .msg email files, raw voice dictations, weirdly formatted spreadsheets.
Miles · 15:34It even runs OCR optical character recognition on physical scanned papers so they become searchable text.
Nora · 15:41But what's truly fascinating is their business model. It perfectly mirrors the shift Andrej Karpathy predicted.
Miles · 15:48Yeah, the fundamental design principle of DocuStrata is to subsidize the ingestion and monetize the interrogation.
Nora · 15:54Subsidize the ingestion.
Miles · 15:55Right. Storing your 100,000 documents is dirt cheap. They practically give that away. way. The actual value, the thing they meter and charge you for, is the act of asking a question and getting an answer. The machine reading the text is treated as a cheap commodity. Your query is the premium product. But looking at their feature list,
Nora · 16:13there is a function called answer memory that seems to fleet dangerously with that knowledge-based poisoning we were just warning about. Oh yeah, big time. When the AI generates a really good, well-cited answer to a complex query, DocuStrata files that generated answer back into your archive as a new searchable document.
Miles · 16:31It is literally saving its own summaries alongside your raw original data.
Nora · 16:38Which is a massive architectural risk.
Miles · 16:40They are prioritizing the speed and low compute cost of retrieving a past answer over the strict data hygiene of forcing a fresh query every time.
Nora · 16:49And there is an even bigger engineering contradiction in this product fact sheet that we have to talk about.
Miles · 16:54A retrieval method.
Nora · 16:55Yeah. To search through those 100,000 documents, DocuStrata uses a technology called ARGI, Retrieval Augmented Generation.
Miles · 17:02Right.
Nora · 17:03Let's explain how that works mechanically. ARGI doesn't read your whole archive at once. It chops all your documents into tiny chunks, turns the text into mathematical coordinates called vectors, and plots them in a high-dimensional space.
Miles · 17:15Sound complicated.
Nora · 17:16It is, but basically when you ask a question, the system finds the text chunks that are mathematically closest to your query, pulls just those few paragraphs and feeds them to the AI to generate an answer.
Miles · 17:26It's an incredibly efficient retrieval mechanism.
Nora · 17:29It is. But wait, Sam Altman and Andrej Karpathy are predicting massive trillion token context windows that just load everything at once.
Miles · 17:39Right. No vector math, no chopping into chunks, no retrieval necessary.
Nora · 17:43This brute force reading the entire archive in one gulp. So is DocuStrata just building elaborate scaffolding around a limitation that is about to disappear? I mean, is R just a technological dead end?
Miles · 17:55Well, this is the billion-dollar gamble happening in the industry right now. Today, R-Ag is what makes reading 100,000 documents economically survivable. Brute-forcing a massive context window for every single question you ask would be financially ruinous with current compute costs.
Nora · 18:11Because it's reading the whole thing every time.
Miles · 18:13Exactly. But beyond just the cost, Aravind Srinivas, the CEO of Perplexity, makes a crucial point with the psychology of the user.
Nora · 18:19Oh, what's that?
Miles · 18:20When humans stop reading full documents, the real, durable value of a product isn't just its ability to generate an answer. The real value becomes verification.
Nora · 18:29Ah, because if you can't verify where the machine got its information, it's just a hallucination engine.
Miles · 18:35Exactly. RAD forces the system to pull specific, discrete chunks of text. which means it can provide the exact highlighted citation back to the original source document.
Nora · 18:47Oh, I see.
Miles · 18:48Yeah, that citation infrastructure. The ability to prove the answer is true without forcing the human to read the whole document. That might end up being the actual feature that retains users long after massive context windows become cheap.
Nora · 19:02Wow. The core paradigm shift here is just profound. I mean, think about your own personal archive, your professional documents, your research. it is no longer a filing cabinet that you painstakingly organize and retrieve things from.
Miles · 19:15No, it's an interface that you interrogate.
Nora · 19:16Exactly. The dead labor of filing, sorting, and skimming is gone. The value is entirely in your ability to ask the right questions.
Miles · 19:24And, you know, that leads to a final thought for you to mull over. Something that builds on all these technical shifts, but points to an even stranger horizon.
Nora · 19:31Lay it on us.
Miles · 19:31Well, we've spent this entire deep dive unpacking what happens when machines read everything for us. But if humans no longer need to read, how long until we no longer need to write?
Nora · 19:42Oh, wow.
Miles · 19:42If 99.9% of content is about to be optimized purely for AI ingestion, we may very soon see an internet populated entirely by dense, ultra-compressed machine languages.
Nora · 19:53Code generated by AI specifically for other AIs to process.
Miles · 19:58Exactly. Bypassing human language entirely. We could easily end up locking ourselves out of our own digital ecosystem, completely unable to comprehend the very internet we created simply because human eyes were no longer the target audience.
Nora · 20:11What a wild thought to end on. We started by imagining the relief of never having to read a dense, confusing manual again. And the logical conclusion might be an internet that is entirely unreadable to us by design.
Miles · 20:22It's a real possibility.
Nora · 20:23Thank you so much for joining us on this deep dive. Keep asking great questions because soon that might be the only job we have left.